- Количество слайдов: 32
Artificial Intelligence and the Internet Edward Brent University of Missouri – Columbia and Idea Works, Inc. Theodore Carnahan Idea Works, Inc.
Overview n n n Objective – Consider how AI can be (and in many cases is being) used to enhance and transform social research on the Internet Framework – intersection of AI and research issues View Internet as a source of data whose size and rate of growth make it important to automate much of the analysis of data
Overview (continued) n n n We discuss a leading AI-based approach, the semantic web, and an alternative paradigmatic approach, and the strengths and weaknesses of each We explore how other AI strategies can be used including intelligent agents, multi-agent systems, expert systems, semantic networks, natural language understanding, genetic algorithms, neural networks, machine learning, and data mining We conclude by considering implications for future research
Key Features of the Internet n n n Decentralized Few or no standards for much of the substantive content Incredibly diverse information Massive and growing rapidly Unstructured data
The Good News About the Internet n n n A massive flow of data Digitized A researcher’s dream
The Bad News n n n A massive flow of data Digitized A researcher’s nightmare
Data Flows n n n The Internet provides many examples of data flows. A data flow is an ongoing flux of new information, often from multiple sources, and typically large in volume. Data flows are the result of ongoing social processes in which information is gathered and/or disseminated by humans for the assessment or consumption by others. Not all data flows are digital, but all flows on the Internet are. Data flows are increasingly available over the internet. Examples of data flows include ¡ ¡ ¡ News articles Published research articles e. Mail Medical records Personnel records Articles submitted for publication Research proposals Arrest records Birth and death records
Data Flows vs Data Sets n Data flows are fundamentally different from the data sets with which most social scientists have traditionally worked. A data set is a collection of data, often collected for a specific purpose and over a specific period of time, then frozen in place. A data flow is an ongoing flux of new information, with no clear end in sight. Data sets typically must be created in research projects funded for that purpose in which relevant data are collected, formatted, cleaned, stored, and analyzed. Data flows are the result of ongoing social processes in which information is gathered and/or disseminated by humans for the assessment or consumption by others. Data sets are sometimes analyzed only once in the context of the initial study, but are often made available in data archives to other researchers for further analysis. Data flows often merit continuing analysis, not only of delimited data sets from specific time periods, but as part of ongoing monitoring and control efforts.
The Need for Automating Analysis n n n Together, the tremendous volume and rate of growth of the Internet, and the prevalence of ongoing data flows make automating analysis both more important and more costeffective. Greater cost savings result from automated analysis with very large data sets Ongoing data flows require continuing analysis and that also makes automation cost-effective
The Semantic Web n n The semantic web is an effort to build into the World Wide Web tags or markers for data along with representations of the semantic meaning of those tags (Berners-Lee and Lassila, 2001; Shadbolt, Hall and Berners-Lee, 2006). The semantic web will make it possible for computer programs to recognize information of a specific type in any of many different locations on the web and to “understand” the semantic meaning of that information well enough to reason about it. This will produce interoperability – the ability of different applications and databases to exchange information and to be able to use that information effectively across applications. Such a web can provide an infrastructure to facilitate and enhance many things including social science research.
Implementing the Semantic Web Contemporary Research Possible Implementation of the Semantic Web Coding scheme XML Schema – a standardized set of XML tags used to markup web pages. For example, research proposals might include tags such as
AI Strategies and the Semantic Web n Several components of the semantic web make use of artificial intelligence (AI) strategies Semantic Web Artificial intelligence and Component related computational strategies Knowledge representation Object-Attribute-Value (O-A-V) triplets commonly used in semantic networks Theory Semantic network Analysis Intelligent agents, Expert systems, Multi-agent models Distributed computing, parallel processing, grid
Strengths of the Semantic Web n Fast and efficient to develop ¡ n Fast and efficient to use ¡ ¡ ¡ n Most coding done by web developers one time and used by everyone Intelligent agents can do most of the work with little human intervention Structure provided makes it easier for computers to process Can take advantage of distributed processing and grid computing Interoperability ¡ Many different applications can access and use information from throughout the web
Weaknesses of the Semantic Web (Pragmatic Concerns) n Seeks to impose standardization on a highly decentralized process of web development ¡ ¡ n Requires cooperation of many if not all developers Imposes the double burden of expressing knowledge for humans and for computers How will tens of millions of legacy web sites be retrofitted? What alternative procedures will be needed for noncompliant web sites? Major forms of data on the web are provided by untrained users unlikely to be able to markup for the semantic web ¡ E. g. , blogs, input to online surveys, emails,
Weaknesses of the Semantic Web (Fundamental Concerns) n Assumes there is a single ontology that can be used for all web pages and all users (at least in some domain). ¡ n n n For example, a standard way to markup products and prices in commercial web sites could make it possible for intelligent agents to search the Internet for the best price for a particular make and model of car. This assumption may be inherently flawed for social research for two reasons. 1) Multiple paradigms - What ontology could code web pages from multiple competing paradigms or world views (Kuhn, 1969). ¡ If reality is socially constructed, and “beauty is in the eye of the beholder” how can a single ontology represent such diverse views? 2) Competing interests – What if developers of web pages have political or economic interests at odds with some of the viewers of those web pages?
Paradigmatic Approach n n n We describe an alternative approach to the semantic web, one that we believe may be more suitable for many social science research applications. Recognizes there may be multiple incompatible views of data Data structure must be imposed on data dynamically by the researcher as part of the research process ¡ (in contrast to the semantic web which seeks to build an infrastructure of web pages with data structure pre-coded by web developers)
Paradigmatic Approach (continued) n n Relies heavily on natural language processing (NLP) strategies to code data. NLP capabilities are not already developed for many of these research areas and must be developed. Those NLP procedures are often developed and refined using machine learning strategies. We will compare the paradigmatic approach to traditional research strategies and the Semantic Web for important research tasks.
Example Areas Illustrating the Paradigmatic Approach n n n Event analysis in international relations Essay grading Tracking news reports on social issues or for clients ¡ n n n E. g. , Campaigns, Corporations, Press agents Each of these areas illustrate significant data flows. These areas and programs within them illustrate elements of the paradigmatic approach. Most do not yet employ all the strategies.
Essay Grading n n These are programs that allow students to submit essays using the computer then a computer program examines the essays and computes a score for the student. Some of the programs also provide feedback to the student to help them improve. These programs are becoming more common for standardized assessment tests and classroom applications. Examples of programs ¡ ¡ ¡ n SAGrader™ E-rater® C-rater® Intelligent Essay Assessor® Criterion® These programs illustrate large ongoing data flows and generally reflect the paradigmatic approach.
Digitizing Data Task Traditional Research Semantic Web Data from Internet digitized Data digitized by by web page developers. web page Other data must be developers Digitizing digitized by researcher or analyzed manually. This can be a huge hurdle. Paradigmatic Approach Data digitized by web page developers The first step in any computer analysis must be converting relevant data to digital form where it is expressed as a stream of digits that can be transmitted and manipulated by computers These two approaches both rely on web page developers to digitize information. This gives them a distinct advantage over traditional research where digitizing data can be a major hurdle.
Essay Grading: Digitizing Data n Digitizing ¡ Papers replaced with digital submissions n ¡ SAGrader, for example, has students submit their papers over the Internet using standard web browsers. Digitizing often still a major hurdle limiting use n n Access issues Security concerns
Data Conversions Task Traditional Research Semantic Web Paradigmatic Approach Converted Data Digitized data suitable for web delivery for human interpretation Digitized data suitable for web delivery and machine interpretation Converting No further data conversions required once digitized by web page author Further conversion sometimes required by researcher (e. g. , OCR, speech recognition, handwriting recognition)
Essay Grading: Converting Data n Data conversion ¡ Where essays are submitted on paper, optical character recognition (OCR) or handwriting recognition programs must be used to convert to digitized text. n Standardized testing programs often face this issue
Encoding Data Task Traditional Research Semantic Web Paradigmatic Approach Coded Data Researchers must encode massive amounts of data Encoding automated using NLP strategies (including statistical, linguistic, rule-based expert systems, and combined strategies) machine learning (unsupervised learning, neural networks, genetic algorithms, data mining) Coded data based on coding rubric Encoding Data Encoding done Each web page by researcher developer must (often with use encode small or of qualitative moderate amount of or quantitative data programs) XML markup based on ontology for that paradigm An XML schema indicates the basic structure expected for a web page XML markup based on standard ontology An XML schema indicates the basic structure expected for a web page
Essay Grading: Coding n n Essay grading programs employ a wide array of strategies for recognizing important features in essays. Intelligent Essay Assessor (IEA) employs a purely statistical approach, latent semantic analysis (LSA). ¡ n E-rater uses a combination of statistical and linguistic approaches. ¡ n It uses syntactic, discourse structure, and content features to predict scores for essays after the program has been trained to match human coders. SAGrader uses a strategy that blends linguistic, statistical, and AI approaches. ¡ n This approach treats essays like a “bag of words” using a matrix of word frequencies by essays and factor analysis to find an underlying semantic space. It then locates each essay in that space and assesses how closely it matches essays with known scores. It uses fuzzy logic to detect key concepts in student papers and a semantic network to represent the semantic information that should be present in good essays. All of these programs require learning before they can be used to grade essays in a specific domain.
Knowledge Task Traditional Research Theory Knowledge Semantic Web A single shared worldview or objective reality Coding scheme Ontology (knowledgebase implemented with a developed by web page Codebook (often developers and shared as imperfect) standard) (implemented with RDF and ontological languages) Paradigmatic Approach Multiple paradigms Multiple ontologies, one for each paradigm (developed by researchers and shared within paradigm) (implemented with RDF and ontological languages)
Essay Grading: Knowledge n n Most essay grading programs have very little in the way of a representation of theory or knowledge. This is probably because they are often designed specifically for grading essays and are not meant to be used for other purposes requiring theory, such as social science research. ¡ n For example, C-rater, a program that emphasizes semantic content in essays, yet has no representation of semantic content other than as desirable features for the essay. The exception is SAGrader. ¡ SAGrader employs technologies developed in a qualitative analysis program, Qualrus. Hence, SAGrader uses a semantic network to explicitly represent and reason about the knowledge or theory.
Analysis Task Analysis Traditional Research Semantic Web Paradigmatic Approach Analysis (by hand, perhaps with help of qualitative or quantitative programs) Intelligent Agents Intelligent agents The semantic web and paradigmatic approaches can take similar approaches to analysis.
Essay Grading: Analysis n n All programs produce scores, though the precision and complexity of the scores varies. Some produce explanations Most of these essay grading programs simply perform a one-time analysis (grading) of papers. However some of them, such as SAGrader, provide for ongoing monitoring of student performance as students revise and resubmit their papers. Since essays presented to the programs are already converted into standard formats and are submitted to a central site for processing, there is no need for the search and retrieval capabilities of intelligent agents
Advantages of Paradigmatic Approach n n Suitable for multiple-paradigm fields Suitable for contested issues Does not require as much infrastructure development on the web Can be used for new views requiring different codes with little lag time
Disadvantages of Paradigmatic Approach n n n Relies heavily on NLP technologies that are still evolving May not be feasible in some or all circumstances Requires extensive machine learning Often requires additional data conversion for automated analysis Requires individual web pages to be coded once for each paradigm rather than a single time, hence increasing costs. (However, by automating this, costs are made manageable) Current NLP capabilities are limited to problems of restricted scope. Instead of general-purpose NLP programs, they are better characterized as special-purpose NLP programs.
Discussion and Conclusions n n n Both semantic web and paradigmatic approaches have advantages and disadvantages Codes on semantic web could facilitate coding by paradigmatic-approach programs Where there is much consensus the single coding for the semantic web could be sufficient While the infrastructure for the semantic web is still in development the paradigmatic approach could facilitate analysis of legacy data The paradigmatic approach could be used to build out the infrastructure for the semantic web